Qwen-Image-Edit Rapid-AIO v23 — headless edit API (tour / MI50)

Runs Phr00t's Rapid-AIO NSFW v23 (the same tune as the gradio app) as a Q8 GGUF on the AMD Instinct MI50 32GB (tour, gfx906, ROCm 5.7, torch 2.3.1), served by ComfyUI behind a thin FastAPI throughput endpoint: image + prompt in → edited PNG out.

Layout on `tour`

/media/tour/APPS/comfyui/
  venv/                       torch 2.3.1+rocm5.7 + ComfyUI deps
  ComfyUI/                    backend + ComfyUI-GGUF custom node
    models/unet/Qwen-Rapid-NSFW-v23_Q8_0.gguf          (21.8 GB)
    models/text_encoders/qwen_2.5_vl_7b_fp8_scaled.safetensors (9.4 GB)
    models/vae/qwen_image_vae.safetensors              (0.25 GB)
  api/                        edit_api.py + workflow + scripts

Run

# 1) backend (terminal 1)
/media/tour/APPS/comfyui/api/run_comfyui.sh
# 2) API (terminal 2)
/media/tour/APPS/comfyui/api/start_api.sh

Use

curl -s -X POST http://192.168.1.160:8500/edit \
  -F image=@car.png \
  -F 'prompt=put the car on a beach at sunset' \
  -F seed=-1 -F steps=4 \
  -o out.png

Form fields: prompt (required), seed (-1 = random), steps (4), cfg (1.0), sampler_name (euler_ancestral), scheduler (beta), max_area (px, 0 = server default ~1MP). Response headers expose X-Seed, X-Width, X-Height.

Settings match Phr00t's v23 recommendation: euler_ancestral / beta, 4 steps, CFG 1. Output size tracks the input aspect ratio, scaled to MAX_AREA.

Performance (MI50, gfx906, 4 steps)

Compute-bound — the 20.8GB unet stays resident; only the text encoder swaps.

Output budget	Latency
0.59 MP	~110 s
0.79 MP (default)	~140 s
1.0 MP	~180 s

Gotchas (why it's built this way)

ComfyUI pinned to v0.3.77. master imports comfy_kitchen, which needs torch.library.custom_op (torch ≥ 2.4). gfx906/ROCm-5.7 tops out at the proven torch 2.3.1+rocm5.7, so newer ComfyUI won't import. v0.3.77 is the newest tag that still has TextEncodeQwenImageEditPlus without that dependency.
--use-split-cross-attention (in run_comfyui.sh) is required: the default pytorch attention path OOMs a 20B edit model on 32GB. Split attention chunks the attention matmul so it fits.
The audio custom-node import errors in the ComfyUI log (libcudart.so.13) are harmless — a CUDA torchaudio got pulled in; image editing doesn't use it.

Manage (local machine)

From this machine, use the deploy_api.sh script:

cd tour-comfy
./deploy_api.sh deploy tour   # sync files and (re)start services
./deploy_api.sh stop tour     # stop and disable services
./deploy_api.sh status tour   # check service status and logs

Manage (on `tour`)

sudo bash /media/tour/APPS/comfyui/api/deploy.sh  # (re)install and start services
sudo bash /media/tour/APPS/comfyui/api/stop.sh    # stop and disable services

Logs

journalctl -u comfyui-backend -f
journalctl -u comfyui-api -f

3.0 KiB Raw Blame History